2.6.17-rc1-mm3 dies in LTP on amd64

2.6.17-rc1-mm3 dies in LTP on amd64

[Martin Bligh]

Runs most tests just fine, but not LTP.
-mm2 ran LTP fine.

Full log here:
http://test.kernel.org/abat/28728/debug/console.log

The trainwreck starts with:

Modules linked in:
Pid: 228, comm: kswapd0 Not tainted 2.6.17-rc1-mm3-autokern1 #1
RIP: 0010:[<ffffffff8047a8dc>] <ffffffff8047a8dc>{__sched_text_start+1852}
RSP: 0000:0000000000000000  EFLAGS: 00010046
RAX: 0000000000000001 RBX: 0000000000000000 RCX: ffffffff805d9338
RDX: ffff8100010c5090 RSI: ffffffff805d9338 RDI: ffff8100010c5090
RBP: ffffffff805d9338 R08: 0000000000000010 R09: ffff8100e3e63d28
R10: ffff8100e3e63a88 R11: 000000000000000b R12: ffff810000011280
R13: ffff81007e186f40 R14: ffff810008003620 R15: 000002b9f81aa1c4
FS:  0000000000000000(0000) GS:ffffffff805fa000(0000) knlGS:00000000f7ea7460
CS:  0010 DS: 0018 ES: 0018 CR0: 000000008005003b
CR2: fffffffffffffff8 CR3: 000000007174a000 CR4: 00000000000006e0
Process kswapd0 (pid: 228, threadinfo ffff8100e3e62000, task 
ffff8100010c5090)
Stack: ffffffff80578e20 ffff8100010c5090 0000000000000001 ffffffff80578f58
        0000000000000000 ffffffff80578e78 ffffffff8020b082 ffffffff80578f58
        0000000000000000 ffffffff80483520
Call Trace: <#DF> <ffffffff8020b082>{show_registers+140}
        <ffffffff8020b30b>{__die+159} <ffffffff8020b380>{die+50}
        <ffffffff8020bb46>{do_double_fault+115} 
<ffffffff8020aa61>{double_fault+125}
        <ffffffff8047a8dc>{__sched_text_start+1852} <EOE>

Code: e8 0d da d8 ff 65 48 8b 34 25 00 00 00 00 4c 8b 46 08 f0 41

Re: 2.6.17-rc1-mm3 dies in LTP on amd64

[Andi Kleen]

On Tuesday 18 April 2006 18:59, Martin Bligh wrote:
> Runs most tests just fine, but not LTP.
> -mm2 ran LTP fine.

I don't think it's my patchkit - currently only has harmless things.

> Full log here:
> http://test.kernel.org/abat/28728/debug/console.log
> 
> The trainwreck starts with:
> 
> Modules linked in:
> Pid: 228, comm: kswapd0 Not tainted 2.6.17-rc1-mm3-autokern1 #1
> RIP: 0010:[<ffffffff8047a8dc>] <ffffffff8047a8dc>{__sched_text_start+1852}
> RSP: 0000:0000000000000000  EFLAGS: 00010046
> RAX: 0000000000000001 RBX: 0000000000000000 RCX: ffffffff805d9338
> RDX: ffff8100010c5090 RSI: ffffffff805d9338 RDI: ffff8100010c5090
> RBP: ffffffff805d9338 R08: 0000000000000010 R09: ffff8100e3e63d28
> R10: ffff8100e3e63a88 R11: 000000000000000b R12: ffff810000011280
> R13: ffff81007e186f40 R14: ffff810008003620 R15: 000002b9f81aa1c4
> FS:  0000000000000000(0000) GS:ffffffff805fa000(0000) knlGS:00000000f7ea7460
> CS:  0010 DS: 0018 ES: 0018 CR0: 000000008005003b
> CR2: fffffffffffffff8 CR3: 000000007174a000 CR4: 00000000000006e0
> Process kswapd0 (pid: 228, threadinfo ffff8100e3e62000, task 
> ffff8100010c5090)
> Stack: ffffffff80578e20 ffff8100010c5090 0000000000000001 ffffffff80578f58
>         0000000000000000 ffffffff80578e78 ffffffff8020b082 ffffffff80578f58
>         0000000000000000 ffffffff80483520
> Call Trace: <#DF> <ffffffff8020b082>{show_registers+140}
>         <ffffffff8020b30b>{__die+159} <ffffffff8020b380>{die+50}
>         <ffffffff8020bb46>{do_double_fault+115} 
> <ffffffff8020aa61>{double_fault+125}
>         <ffffffff8047a8dc>{__sched_text_start+1852} <EOE>


Not very useful.  Something double faulted, but it's not on the stack
[I wonder if the stack walker over double faults is broken. Jan - did you
ever test that after you redid the walker?]

If you can reproduce it on a Intel machine it might be possible to find
it using the last branch registers (patch for that available on request).

Otherwise binary search I guess.

-Andi

Re: 2.6.17-rc1-mm3 dies in LTP on amd64

[Jan Beulich]

>Not very useful.  Something double faulted, but it's not on the stack
>[I wonder if the stack walker over double faults is broken. Jan - did you
>ever test that after you redid the walker?]

Maybe not specifically for a double fault, but I'm certain I checked it for some of the IST ones; looking at the code I
also can't see anyting obviously wrong. But clearly, a double fault has its most likely reason being a bad kernel stack
pointer prior to a page (or other) fault. It would therefore be necessary to know the value of the stack pointer as
retrieved from the double fault stack, which with the current display logic is not possible.

Jan